Models
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresNot ranked · too few of its benchmarks are in the capability model to estimate a score
MAI Code 2 is Microsoft AI's larger agentic coding model, built for long-horizon software engineering in real repositories and released on November 17, 2026 in GitHub Copilot. Its model card describes a sparse mixture-of-experts transformer with an integrated vision encoder, pre-trained on roughly 50 trillion tokens and post-trained by Microsoft from scratch, with reinforcement learning run in the GitHub Copilot CLI and VS Code agent harnesses; the card's summary gives 80B active parameters, while its overview gives 77 billion active and about 2.8 trillion total. It accepts text and image input (922,000 input tokens) and returns text (128,000 output tokens) in a 1,050,000-token context, with low, medium, high, xhigh and max reasoning effort. Weights are not distributed, and Microsoft gives no per-token price, showing launch costs at a 50% promotional rate.
Scored on 4 benchmarks · 4 total scores · 0 independently reported
Organizations reporting headline scores for 4 benchmarks
| Software · 3 benchmarks | |||||
|---|---|---|---|---|---|
| DeepSWE v1.1In fit | 16 / 81 | ~3 mo | ~Apr 2027 | 2026-11-17 | |
| CyberGymIn fitnot reported | 27 / 50 | ~4 mo | Jul 2026 | 2026-11-17 | |
| Terminal-Bench 2.1In fit | 31 / 226 | ~4 mo | Nov 2026* | 2026-11-17 | |
| Science · 1 benchmark | |||||
| GPQA DiamondIn fitnot reported | 34 / 415 | ~5 mo | Apr 2026 | 2026-11-17 | |